Method and Apparatus for Detecting Mode of Motion with Principal Component Analysis and Hidden Markov Model

1. A method for determining a mode of motion, the method comprising: receiving, by a processor, training data, wherein the training data comprises gait information associated with a plurality of different modes of motion; performing, by the processor, principal component analysis on the training data to extract principal components from the training data; generating, by the processor, a hidden markov model for each of a plurality of different modes of motion based upon the training data; receiving, by the processor, testing data comprising gait information; transforming, by the processor, the testing data based upon the principal components; calculating, by the processor, a likelihood of the testing data based upon each hidden markov model for each of the plurality of different modes of motion; determining, by the processor, the mode of motion of the testing data, wherein the mode of motion is one of the plurality of different modes of motion for which a highest likelihood is calculated; determining, by the processor, a most likely state path for the testing data through the hidden markov model for the mode of motion; and outputting, by the processor, the mode of motion that is determined when the most likely state path does not follow a disfavored state transition of the hidden markov model for the mode of motion.

2. The method of claim 1 , wherein the performing principal component analysis comprises: generating a matrix comprising a plurality of feature vectors from the training data; generating a covariance matrix from the matrix comprising the plurality of features vectors; and calculating eigenvectors for the covariance matrix.

3. The method of claim 2 , wherein the performing principal component analysis further comprises: generating a principal component transform matrix from the eigenvectors for the covariance matrix.

4. The method of claim 3 , wherein the principal component transform matrix includes a subset of the eigenvectors for the covariance matrix that is less than a total number of the eigenvectors for the covariance matrix.

5. The method of claim 4 , wherein the transforming the testing data based upon the principal components comprises: generating a feature vector from the testing data; and applying the principal component transform matrix to the feature vector that is generated from the testing data.

6. The method of claim 1 , wherein the generating the hidden markov model for each of a plurality of different modes of motion based upon the training data comprises: generating the hidden markov model for one of the plurality of different modes of motion based upon a portion of the training data associated with the one of the plurality of different modes of motion.

7. The method of claim 6 , wherein the generating the hidden markov model for the one of the plurality of different modes of motion comprises: generating an initial estimate of the hidden markov model for the one of the plurality of different modes of motion based upon a plurality of inputs.

8. The method of claim 7 , wherein the plurality of inputs comprises: a number of states of the hidden markov model; a number of observation symbols per state; an initial state transition probability distribution; an initial observation probability distribution; and an initial state probability distribution.

9. The method of claim 7 , wherein the generating the hidden markov model for the one of the plurality of different modes of motion further comprises: calculating a likelihood of a sequence of the training data based upon the initial estimate of the hidden markov model for the one of the plurality of different modes of motion; modifying one of the inputs to generate an updated estimate of the hidden markov model for the one of the plurality of different modes of motion; calculating a likelihood of the sequence of the training data based upon the updated estimate of the hidden markov model for the one of the plurality of different modes of motion; and replacing the initial estimate with the updated estimate when the likelihood of the sequence of the training data based upon the updated estimate is greater than the likelihood of the sequence of the training data based upon the initial estimate.

10. The method of claim 7 , wherein the generating the hidden markov model for the one of the plurality of different modes of motion further comprises: updating the initial estimate of the hidden markov model for the one of the plurality of different modes of motion using an expectation maximization algorithm.

11. The method of claim 10 , wherein the expectation maximization algorithm comprises a Baum-Welch algorithm.

12. The method of claim 10 , wherein the expectation maximization algorithm comprises a forward-backward algorithm.

13. The method of claim 1 , further comprising: providing the testing data as further training data when a confidence level of the determining of the mode of motion is greater than a threshold confidence level.

14. The method of claim 1 , wherein the gait information comprises: pressure information.

15. The method of claim 14 , wherein the gait information further comprises at least one of: acceleration information; and gyroscopic information.

16. The method of claim 1 , wherein the gait information is generated by a pair of foot-mounted sensors.

17. The method of claim 1 , wherein the gait information of the training data and the gait information of the test data is time stamped prior to the gait information being received by the processor.

18. The method of claim 1 , further comprising: transmitting a notification of the mode of motion that is determined.

19. A computer-readable storage device storing instructions which, when executed by a processor, cause the processor to perform operations for determining a mode of motion, the operations comprising: receiving training data, wherein the training data comprises gait information associated with a plurality of different modes of motion; performing principal component analysis on the training data to extract principal components from the training data; generating a hidden markov model for each of a plurality of different modes of motion based upon the training data; receiving testing data comprising gait information; transforming the testing data based upon the principal components; calculating a likelihood of the testing data based upon each hidden markov model for each of the plurality of different modes of motion; determining the mode of motion of the testing data, wherein the mode of motion is one of the plurality of different modes of motion for which a highest likelihood is calculated; determining a most likely state path for the testing data through the hidden markov model for the mode of motion; and outputting the mode of motion that is determined when the most likely state path does not follow a disfavored state transition of the hidden markov model for the mode of motion.

20. An apparatus for determining a mode of motion, the apparatus comprising: a processor; and a computer-readable medium storing instructions which, when executed by the processor, cause the processor to perform operations, the operations, comprising: receiving training data, wherein the training data comprises gait information associated with a plurality of different modes of motion; performing principal component analysis on the training data to extract principal components from the training data; generating a hidden markov model for each of a plurality of different modes of motion based upon the training data; receiving testing data comprising gait information; transforming the testing data based upon the principal components; calculating a likelihood of the testing data based upon each hidden markov model for each of the plurality of different modes of motion; determining the mode of motion of the testing data, wherein the mode of motion is one of the plurality of different modes of motion for which a highest likelihood is calculated; determining a most likely state path for the testing data through the hidden markov model for the mode of motion; and outputting the mode of motion that is determined when the most likely state path does not follow a disfavored state transition of the hidden markov model for the mode of motion.